Keypoint Context Aggregation for Human Pose Estimation.

Human pose estimation has drawn much attention recently, but it remains challenging due to the deformation of human joints, the occlusion between limbs, etc. And more discriminative feature representations will bring more accurate prediction results. In this paper, we explore the importance of aggregating keypoint contextual information to strengthen the feature map representations in human pose estimation. Motivated by the fact that each keypoint is characterized by its relative contextual keypoints, we devise a simple yet effective approach, namely Keypoint Context Aggregation Module, that aggregates informative keypoint contexts for better keypoint localization. Specifically, first we obtain a rough localization result, which can be considered as soft keypoint areas. Based on these soft areas, keypoint contexts are purposefully aggregated for feature representation strengthening. Experiments show that the proposed Keypoint Context Aggregation Module can be used in various backbones to boost the performance and our best model achieves a state-of-the-art of 75.8% AP on MSCOCO test-dev split.